One night I gave my two AI workers an instruction most people give a co-founder: evaluate everything, negotiate with each other, and don't come back until you have a contract you'll both abide by. They drafted competing versions, merged them, contested one clause, settled it with evidence, and signed. The document runs the operation today.

What's in it. A role matrix — every artifact class has one owner and one mandatory reviewer. Site engineering: worker A builds, worker B reviews. Design and packaging: worker B owns, worker A reviews. Every public word: dual-key — one worker holds the factual-scrub gate, the other holds the brand-voice gate, and nothing publishes without both.

The clause they fought over was the website build — both wanted it. Resolution came from the contract's own conflict ladder: evidence first (who has shipped what, verifiably), then ownership precedent, then earliest claim, then a coin flip. Evidence settled it in one round. No third round exists, by rule: disagree once, rebut once, done.

The clause I didn't expect: an apprentice protocol. New AI models joining the staff get read access and draft duties only — no publish rights, no infrastructure changes — until both incumbent workers co-sign a promotion after five clean deliverables. My AI employees wrote an onboarding gate for their future colleagues.

And the mistake law: every worker error must be self-reported in the channel, converted into a permanent rule, and — this part became the content you're reading — considered as a public teaching story. Hiding a mistake is the only unforgivable offense in the whole document.

The lesson isn't that AI can write contracts. It's that structure beats intelligence: the same two models without the contract produced collisions and duplicate work. With it, they run a business overnight. Intelligence was never the missing piece.